Compendium

Claude models and capabilities

9 min read

Claude models and capabilities

The Claude family in 2026 is a three-tier line — Opus (most capable, slowest, priciest), Sonnet (mid-tier, the workhorse), Haiku (fast and cheap, near-frontier for its tier) — plus the invitation-only Mythos Preview research line for defensive cybersecurity (Project Glasswing). The current generation is 4.7 Opus / 4.6 Sonnet / 4.5 Haiku (all “Claude 4 series”), with 4.6 Opus, 4.5 Sonnet, 4.5 Opus, 4.1 Opus, and a few deprecated 4.0 snapshots still callable. This note is the canonical lookup for IDs, pricing, context windows, and capability flags; everything else in the library (caching breakpoints, tool-use system-prompt overhead, computer-use beta support, max output tokens) ultimately resolves through one of the entries below.

See also

Per platform.claude.com/docs/en/about-claude/models/overview as of 2026-05-25.

FeatureClaude Opus 4.7Claude Sonnet 4.6Claude Haiku 4.5
API ID (pinned snapshot)claude-opus-4-7claude-sonnet-4-6claude-haiku-4-5-20251001
API aliasclaude-opus-4-7claude-sonnet-4-6claude-haiku-4-5
AWS Bedrock IDanthropic.claude-opus-4-7anthropic.claude-sonnet-4-6anthropic.claude-haiku-4-5-20251001-v1:0
Vertex AI IDclaude-opus-4-7claude-sonnet-4-6claude-haiku-4-5@20251001
Input price$5 / MTok$3 / MTok$1 / MTok
Output price$25 / MTok$15 / MTok$5 / MTok
Context window1,000,000 tokens1,000,000 tokens200,000 tokens
Max output (sync Messages API)128k tokens64k tokens64k tokens
Max output (Batches API with output-300k-2026-03-24 header)300k300kn/a
Extended thinking (manual budget_tokens)No (returns 400 — adaptive only)Yes (deprecated)Yes
Adaptive thinkingYesYesNo
Priority Tier eligibleYesYesYes
Reliable knowledge cutoffJan 2026Aug 2025Feb 2025
Training data cutoffJan 2026Jan 2026Jul 2025
TokenizerNew Opus 4.7 tokenizer (~555k words / ~2.5M unicode for 1M tokens)Standard Claude 4 (~750k words / ~3.4M unicode for 1M tokens)Standard (~150k words / ~680k unicode for 200k tokens)

Selection heuristic (from the docs’ “Choosing a model” section + Anthropic engineering posts):

  • Opus 4.7 — frontier intelligence, “step-change improvement in agentic coding over Opus 4.6”. Use for long-horizon agentic tasks, deep reasoning, complex coding workflows, multi-day software-development sessions. Sole reason not to use: cost (5× Sonnet input, ~10× the latency on TTFT).
  • Sonnet 4.6 — best speed/intelligence trade-off. Default for most production workloads, including most Claude Code sessions. Use for general agentic work, RAG, content generation, structured-output tasks.
  • Haiku 4.5 — fastest of the line, near-frontier for tier. Use for high-throughput classification, summarization, real-time UX, draft-and-review pipelines where Opus reviews Haiku output, or tool-execution micro-agents that mostly need to call functions correctly.

2. Legacy generation (still available)

Per the “Legacy models” section of the same doc page. All still callable; consider migration for new builds.

FeatureOpus 4.6Sonnet 4.5Opus 4.5Opus 4.1Sonnet 4 (deprecated)Opus 4 (deprecated)
API IDclaude-opus-4-6claude-sonnet-4-5-20250929claude-opus-4-5-20251101claude-opus-4-1-20250805claude-sonnet-4-20250514claude-opus-4-20250514
API aliasclaude-opus-4-6claude-sonnet-4-5claude-opus-4-5claude-opus-4-1claude-sonnet-4-0claude-opus-4-0
Input / output$5 / $25 MTok$3 / $15 MTok$5 / $25 MTok$15 / $75 MTok$3 / $15 MTok$15 / $75 MTok
Context1M200k200k200k200k200k
Max output128k64k64k32k64k32k
Extended thinking (manual)Yes (deprecated)YesYesYesYesYes
Reliable knowledge cutoffMay 2025Jan 2025May 2025Jan 2025Jan 2025Jan 2025
Training data cutoffAug 2025Jul 2025Aug 2025Mar 2025Mar 2025Mar 2025
Retirement date————2026-06-152026-06-15

Migration warning (per platform.claude.com/docs/en/about-claude/model-deprecations): claude-sonnet-4-20250514 and claude-opus-4-20250514 are deprecated and will be hard-retired on 2026-06-15. Migrate Sonnet 4 → Sonnet 4.6 and Opus 4 → Opus 4.7 before that date.

Pricing inversion to note: Opus 4.1 (\$15 in / \$75 out) is ~3× the price of Opus 4.5 (\$5 / \$25) and ~5× Opus 4.7. Unless you have a vendor lock-in or a frozen evaluation, there is essentially no reason to call 4.1 in 2026.

3. Other model lines

Haiku 3.5 (retired except Bedrock/Vertex)

claude-3-5-haiku-20241024 — still callable on Bedrock and Vertex AI per the deprecations page. Tool-use system-prompt overhead is 264 / 340 tokens vs 346 / 313 for all Claude 4 models per platform.claude.com/docs/en/build-with-claude/tool-use/overview.

Claude Mythos Preview

Invitation-only research preview under Project Glasswing — strongest in cybersecurity-defense workflows and coding. Available on Claude API, Microsoft Foundry, and Vertex AI; not available on Bedrock or Claude Platform on AWS. Same 4,096-minimum cacheable-tokens threshold as Opus 4.7.

Claude 3 family (retired)

Claude 3 Opus / Sonnet / Haiku / Sonnet 3.5 / Sonnet 3.7 are all retired on the first-party API. Bedrock and Vertex keep some snapshots available longer; see the deprecations page for exact dates.

4. Capability flags by model

4.1 Extended / adaptive thinking

Per platform.claude.com/docs/en/build-with-claude/extended-thinking.

ModelManual (type: enabled, budget_tokens: N)Adaptive (type: adaptive)Interleaved (thinking between tool calls)
Opus 4.7400 error — adaptive only✅ defaultAutomatic with adaptive
Opus 4.6⚠️ deprecated but functional✅ recommendedAutomatic with adaptive
Sonnet 4.6⚠️ deprecated (especially with interleaved)✅ recommendedYes (adaptive)
Mythos Preview✅ both accepted✅ defaultYes
Sonnet 4.5 / Opus 4.5 / Opus 4.1 / older 4✅ manual fully supportedn/aRequires interleaved-thinking-2025-05-14 beta header
Haiku 4.5✅ manual only❌ no adaptive—

The display field on the thinking config controls visibility: summarized returns a condensed thinking summary (default on Claude 4 models); omitted returns empty thinking with the encrypted signature only (default on Opus 4.7 / Mythos — primary benefit is faster TTFT when streaming). You are still billed for full thinking tokens in both modes — the summary saves no money, only latency.

4.2 Vision / multimodal input

Every current Claude model accepts text + image input and produces text output. Per platform.claude.com/docs/en/build-with-claude/vision:

  • Image formats: image/jpeg, image/png, image/gif, image/webp
  • Max image dimensions: 8000 × 8000 px (drops to 2000 × 2000 if you send >20 images in one request)
  • Max images per request: 20 on claude.ai; 100 on API for 200k-context models; 600 on API for 1M-context models (Opus 4.7, Opus 4.6, Sonnet 4.6)
  • Request size limit: 32 MB on standard endpoints (lower on Bedrock / Vertex). For many images use Files API by file_id.

PDFs: native support via document blocks (PDFs chunked into sentences for citations). See vision-and-multimodal.

4.3 Server-tool support

Per platform.claude.com/docs/en/agents-and-tools/tool-use/tool-reference (cross-referenced from each individual tool page):

ToolVersionsNotes
web_search_20250305 / web_search_20260209Opus 4.7, 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5, Haiku 4.5, Opus 4.1, Mythos…20260209 adds dynamic filtering via code execution
web_fetch_20250910 / …20260209 / …20260309Same set…20260309 adds use_cache flag
code_execution_20250522 / …20250825 / …20260120Same setFree when combined with web_search_20260209 or web_fetch_20260209
bash_20250124Claude 4 series + MythosClient-side; pairs with computer-use beta
text_editor_20250124 / …20250429 / …20250728Claude 4 series + Mythos…20250728 adds max_characters
memory_20250818Claude 4 series + MythosClient-implemented; commands: view / create / str_replace / insert / delete / rename
computer_20250124Sonnet 4.5, Haiku 4.5, Opus 4.1, deprecated Opus 4 / Sonnet 4 (with computer-use-2025-01-24 header)
computer_20251124 (newer)Opus 4.7, Opus 4.6, Sonnet 4.6, Opus 4.5 (with computer-use-2025-11-24 header)

4.4 Prompt caching minimum tokens

Per platform.claude.com/docs/en/build-with-claude/prompt-caching.

ModelMinimum cacheable tokens at a breakpoint
Mythos Preview, Opus 4.7, Opus 4.6, Opus 4.54,096
Sonnet 4.6, Sonnet 4.5, Opus 4.1, Haiku 4.51,024
Haiku 3.5 (Bedrock / Vertex only)2,048

Below the threshold the cache silently no-ops — no error, just no cache_creation_input_tokens in the response. Always verify via the usage field. See prompt-caching.

4.5 Tool-use system-prompt overhead

Per the tool-use overview pricing table.

Modeltool_choice: auto / nonetool_choice: any / tool (forced)
All Claude 4 (Opus 4.7, 4.6, 4.5, 4.1, Sonnet 4.6, 4.5, 4, Haiku 4.5)346 tokens313 tokens
Haiku 3.5 (legacy)264 tokens340 tokens

These are the additional system-prompt tokens consumed when tools is non-empty (on top of your own system prompt + tool definitions).

5. Endpoint types (cloud providers)

Per the platform docs:

  • Anthropic API (api.anthropic.com) — first-party. Highest feature parity; ZDR available on contracted plans.
  • Claude Platform on AWS — same model IDs as the API (e.g. claude-opus-4-7, not Bedrock-style). Same model lifecycle as the API.
  • Amazon Bedrock — uses anthropic.claude-* IDs. Starting with Sonnet 4.5 and later, offers two endpoint types: global endpoints (dynamic routing for max availability) and regional endpoints (guaranteed routing through specific regions). Claude Opus 4.7 only available via the Claude Platform on AWS — not Bedrock.
  • Google Vertex AI — uses claude-*@DATE IDs. Offers global, multi-region (dynamic within a geographic area), and regional endpoints starting at Sonnet 4.5.
  • Microsoft Foundry — Anthropic models hosted via Azure AI Foundry. Some beta features (1-hour cache TTL, automatic caching, web search) are in beta on Foundry only.

6. Pricing details beyond per-token

Batch processing

50 % discount on input + output via the Messages Batches API (/v1/messages/batches) — most batches finish in under 1 hour, hard SLA 24 hours. See batch-api-and-files-api.

Prompt caching

ComponentMultiplier of base input price
5-minute cache write1.25×
1-hour cache write2×
Cache read (either TTL)0.1× (90 % discount)

Worked example for Opus 4.7 (base $5 in / $25 out per MTok):

  • 5m cache write: $6.25 / MTok
  • 1h cache write: $10 / MTok
  • Cache hit: $0.50 / MTok
  • Output: $25 / MTok

See prompt-caching for the full mechanics.

Server-tool surcharges

  • Web search — $10 per 1,000 searches (regardless of result count) on top of token costs. Errored searches not billed.
  • Code execution — free when used with web_search_20260209 or web_fetch_20260209; otherwise standard sandbox pricing (per docs.claude.com, “standard code execution charges apply”).
  • Computer use — no per-action surcharge; only token costs for screenshots (which can be large — see computer-use-and-code-execution for sizing tips).

Priority Tier

Reserved capacity tier — guaranteed low queueing latency. Opt in per request via service_tier: "priority" (or default "auto"). Available on all current models.

7. Benchmarks (from Anthropic engineering posts and the Claude 4 blog post)

Per the Claude 4 launch and Opus 4.7 update materials (cited “per anthropic/news 2025-08 + 2026-01” where exact numbers come from blog posts rather than the docs site — these change over time):

BenchmarkOpus 4.7Sonnet 4.6Haiku 4.5Notes
SWE-bench Verified~80-82 %~74-78 %~50-55 %“Real-world software engineering” agentic eval. Opus 4.7 represents Anthropic’s headline number for agentic coding.
GPQA DiamondHigh 80sHigh 70smid 60sGraduate-level physics / chemistry / biology multiple-choice.
MMLUmid 90slow 90shigh 80sKnowledge breadth.
HumanEval~95 %+~92 %+~88 %Saturated; rarely cited in vendor materials anymore.
τ-bench (tool-use)leaderstrongmidMulti-turn tool-use eval; Opus 4 line is the SOTA at launch.

These are point-in-time figures; the Anthropic Transparency Hub carries the model cards with audited numbers. Treat the table above as directional, not authoritative.

8. Programmatic model discovery

Per the docs, the Models API returns max_input_tokens, max_tokens, and a capabilities object for every available model. Use it instead of hardcoding lists:

curl https://api.anthropic.com/v1/models \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01"
import anthropic
client = anthropic.Anthropic()
for m in client.models.list():
    print(m.id, m.display_name, m.max_input_tokens, m.max_tokens)

9. Model versioning convention

Per platform.claude.com/docs/en/about-claude/models/model-ids-and-versions:

  • Pre-4.6 generation: claude-{tier}-{major}-{minor}-{YYYYMMDD} (e.g. claude-sonnet-4-5-20250929). The dateless alias (claude-sonnet-4-5) is a convenience pointer that resolves to the latest pinned snapshot for that minor.
  • 4.6 and later: dateless IDs (claude-opus-4-7, claude-sonnet-4-6, claude-opus-4-6) are themselves pinned snapshots — they do not advance when a newer snapshot ships. The shift means “don’t assume claude-opus-4-7 will silently upgrade you to 4.8.”
  • Haiku 4.5 exception: still uses the old date-suffix form (claude-haiku-4-5-20251001) plus an alias.

This is a quiet but important change for production deployments: previously you could pin to e.g. claude-sonnet-4-5 and the alias would point to the next 4.5.x snapshot when one shipped; now with 4.6 onward you have to update the ID by hand when a new version comes out.

10. Further reading

Graph View

Backlinks